RoPE Attention Can Be Trained in Almost Linear Time

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Cao, Yang, Huo, Jiayan, Liang, Yingyu, Shi, Zhenmei, Song, Zhao
Natura: Preprint
Pubblicazione: 2024
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866908783830106112
author Cao, Yang
Huo, Jiayan
Liang, Yingyu
Shi, Zhenmei
Song, Zhao
author_facet Cao, Yang
Huo, Jiayan
Liang, Yingyu
Shi, Zhenmei
Song, Zhao
contents The Rotary Position Embedding (RoPE) mechanism has become a powerful enhancement to the Transformer architecture, which enables models to capture token relationships when encoding positional information. However, the RoPE mechanisms make the computations of attention mechanisms more complicated, which makes efficient algorithms challenging. Earlier research introduced almost linear time algorithms for the forward computation under specific parameter settings of bounded entries (i.e., in time $n^{1+o(1)}$ where $n$ is the number of input tokens), but has not addressed backward computation. In this work, we develop the first almost linear time algorithm for backward computations in the RoPE-based attention under bounded entries. Our approach builds on recent advancements in fast RoPE attention computations, utilizing a novel combination of the polynomial method and the Fast Fourier Transform. Furthermore, we show that with lower bounds derived from the Strong Exponential Time Hypothesis (SETH), the bounded entry condition is necessary for subquadratic performance.
format Preprint
id arxiv_https___arxiv_org_abs_2412_17316
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle RoPE Attention Can Be Trained in Almost Linear Time
Cao, Yang
Huo, Jiayan
Liang, Yingyu
Shi, Zhenmei
Song, Zhao
Machine Learning
Artificial Intelligence
Computational Complexity
Computation and Language
The Rotary Position Embedding (RoPE) mechanism has become a powerful enhancement to the Transformer architecture, which enables models to capture token relationships when encoding positional information. However, the RoPE mechanisms make the computations of attention mechanisms more complicated, which makes efficient algorithms challenging. Earlier research introduced almost linear time algorithms for the forward computation under specific parameter settings of bounded entries (i.e., in time $n^{1+o(1)}$ where $n$ is the number of input tokens), but has not addressed backward computation. In this work, we develop the first almost linear time algorithm for backward computations in the RoPE-based attention under bounded entries. Our approach builds on recent advancements in fast RoPE attention computations, utilizing a novel combination of the polynomial method and the Fast Fourier Transform. Furthermore, we show that with lower bounds derived from the Strong Exponential Time Hypothesis (SETH), the bounded entry condition is necessary for subquadratic performance.
title RoPE Attention Can Be Trained in Almost Linear Time
topic Machine Learning
Artificial Intelligence
Computational Complexity
Computation and Language
url https://arxiv.org/abs/2412.17316